Nigerian Pidgin (Naija) Speech Recognition
Orinode's Pidgin ASR is one of the strongest results in Maraba v1: 9.8% WER on 200 spontaneous samples (May 2026). Nigerian Pidgin — locally Naija — is the most widely spoken Nigerian language with an estimated 75+ million speakers, yet it has no dedicated language slot in any global ASR model. We made one.
What Nigerian Pidgin is
Nigerian Pidgin (ISO 639-3 pcm) is a creole that emerged from English contact with West African languages — predominantly Yorùbá, Igbo, Hausa, Edo, and Efik. It is grammatically and phonologically distinct from any variant of English:
- Verb tense particles — dey (continuous), don (perfective), go (future) replace English morphology.
- Pronoun system — I, you, im/am, we/wuna, dem with collapsed case distinctions.
- Substrate phonology — vowel inventory follows the speaker's L1 (Yorùbá-Pidgin vs Igbo-Pidgin vs Hausa-Pidgin sound noticeably different).
- Lexical borrowing — actively borrows from Yorùbá (wahala), Igbo (nna), Hausa, and slang sources.
Why standard ASR fails on Pidgin
Whisper, Google Speech-to-Text, and Azure all classify Pidgin audio as "broken English" and transcribe it phonetically against an English lexicon — producing output like "we are going to the market" for we dey go market. Worse, they often refuse to transcribe at all when language detection mislabels the segment.
Orinode's approach
- Token-slot remapping: Whisper-large-v3 has no Pidgin language token. We repurposed the rarely-used Hawaiian (
haw) slot — it has minimal real-world Hawaiian audio competing for the signal, so retraining cleanly redirects the slot. - Training data: Mozilla Common Voice 17 Pidgin partition + Orinode-curated Naija news broadcasts + customer-service call recordings (consented and anonymized).
- Code-switch handling: Pidgin sentences frequently switch to Yorùbá / Igbo / Hausa or even formal English. Orinode's training data explicitly models these switches at the utterance level.
Performance (May 2026)
| Metric | Value | N |
|---|---|---|
| WER (normalized) | 9.76% | 200 |
| WER (raw, case-sensitive) | 9.90% | 200 |
Use cases
Pidgin ASR matters most for: customer-service call centers (BVN verification, banking complaints, telco issues), broadcast captioning for Naija-language radio/TV stations, and government services trying to be linguistically inclusive for the majority of Nigerians who use Pidgin as their working everyday language.
Get the model
Maraba v1 weights with the Pidgin (Hawaiian-slot) configuration: huggingface.co/Orinode. Production API: [email protected].
For the deployable voice agent built on this model, see Maraba at maraba.ai.